Slashdot Mirror
DeepMind AI AlphaStar Wins 10-1 Against 'StarCarft II' Pros (newscientist.com)
In a series of matches streamed on YouTube and Twitch, DeepMind AI AlphaStar defeated two top-ranked professionals 10-1 at real-time strategy game StarCraft II. "This is of course an exciting moment for us," said David Silver at DeepMind in a live stream watched by more than 55,000 people. "For the first time we saw an AI that was able to defeat a professional player." New Scientist reports: DeepMind created five versions of their AI, called AlphaStar, and trained them on footage of human games. The different AIs then played against each other in a league, with the leading AI accumulating the equivalent of 200 years of game experience. With this, AlphaStar beat professional players Dario Wunsch and Grzegorz Komincz -- ranked 44th and 13th in the world respectively. AlphaStar's success came with some caveats: the AI played only on a single map, and using a single kind of player (there are three in the game). The professionals also had to contend with playing different versions of AlphaStar from match to match. While the AlphaStar was playing on a single graphics processing unit, a computer chip found in many gaming computers, it was trained on 16 tensor processing units hosted in the Google cloud -- processing power beyond the realms of many.
Actually, the AI is restricted to the viewport, it is using an interface into the game, so no it doesn't have maphack, like a normal scripted AI would.
It's already operating with imperfect knowledge. From Vox:
During the 10 matches, the AI had one big advantage that a human player doesn’t have: It was able to see all of the parts of the map where it had visibility, while a human player has to manipulate the camera.
Emphasis mine. Yes, it's an advantage, but it's not cheating. Humans can use the minimap to see what's going on as well.
The problem with these matches and Starcraft in general, is that the it's able to win just with good micro. So getting really good at Starcraft doesn't get us closer to actual AI.
Wrong for the live game against MaNa. Quote: "Following the broadcast of the recorded matches, DeepMind introduced a new version of AlphaStar that MaNa took on in a live match. The agent that played the live game didn't have the benefit of the overhead camera and instead had to make decisions on where to place its focus in the same way a human would."
But you are right for the recorded games from December 2018.
How can it not get closer? AI didn't used to be able to beat professional StarCraft players. Now they can. That is, by definition, moving closer, because it's not moving backwards, and clearly an improvement from not moving at all.
This is just more goalpost shifting, finding nonsense reasons to argue why it "doesn't count". Consider the alternative that maybe the things humans do aren't as as clever as we tell ourselves. If it's just "good micro", why can't humans use "good micro" to beat the AI, if we're so great?
Those who do not learn from commit history are doomed to regress it.
How can it not get closer? AI didn't used to be able to beat professional StarCraft players. Now they can. That is, by definition, moving closer, because it's not moving backwards, and clearly an improvement from not moving at all.
It is not moving at all. There's no viable route from this to general purpose AI. By playing countless games against itself on the same map, it is still performing a search on a decision tree, weighing the (now more fuzzy) nodes. It did not acquire any actual understanding of Starcraft mechanics. It does not logically reason about anything that it hasn't seen before. If you give it Warcraft instead, it'll take another several months of work from a team of very intelligent humans to make it good at it. In fact, I'll bet a big enough balance patch will cause it to have to throw out everything it's learned.
This is just more goalpost shifting, finding nonsense reasons to argue why it "doesn't count". Consider the alternative that maybe the things humans do aren't as as clever as we tell ourselves.
The goal post has always been to replace human intelligence. I don't see any AI building Starcraft-playing AI's, or discussing how long it will be before they are replaced by even better AIs.
If it's just "good micro", why can't humans use "good micro" to beat the AI, if we're so great?
Because humans have muscles that take time to move?
Playing a multiplayer game with bots used to be seen as an inferior experience to playing with real humans. Now imagine that instead of something like AlphaStar's utility function being set to trying to win, it's set to trying to make the human opponents have the most fun. Of course it'd need some understanding of the mindset of the player; they might not want to always win, or always have close matches, or possibly they're a sore loser. However, this could be inferred somewhat by player behavior (even outside of the match proper, e.g. in menus).
Put that in a game and ship it, and that could be a killer feature. People might prefer to play with a bot that'll guarantee a fun time, over a human that might rage quit or be an unfair match that leads to a one-sided game.
#MakeGamesSinglePlayerAgain
Corruption is convincing someone that the selfless ideal is the same as their selfish ideal.
Are you saying it doesn't?
In competitive StarCraft, you take into consider build orders where you need to have done a specific set of steps in a specific amount of time. If you fail to do it, it puts you behind the opponent and will cause you to lose the game.
It's so meticulous that seasoned casters can actually tell a build based on how many workers a player has in a minute.
Just look at the Liquipedia entry for Protoss build orders and you'll see what I mean. And that's just one race out of three.
It did not acquire any actual understanding of Starcraft mechanics
If it beat professional human players, then yes it did acquire an actual understanding of Starcraft mechanics. In fact, a better understanding than humans.
That you can train a dog to bark twice when shown 1+1, and three times when shown 2+1 does not mean the dog can do arithmetic.